Heterogeneity in meta-analysis of FDG-PET studies to diagnose lung cancer.

Heterogeneity inMeta-analysis of FDG-PET Studies to Diagnose Lung Cancer To the Editor Dr Deppen and colleagues1 conducted a large meta-analysis that showed the limitations of lung cancer diagnosis using fludeoxyglucose F 18 combined with positron emission tomography (FDG-PET) in areas with endemic infectious lung disease. Although the sensitivity and specificity of FDG-PET diagnosis was heterogeneous across the included studies, thereby compromising interpretation of the pooled results, the relevance of presenting an I2 statistic to underscore and interpret the extent of the heterogeneity should be questioned. The large reported I2 values may be an artifact of their chosen measure (proportions) and may not solely reflect important clinical or contextual sources of heterogeneity. The I2 statistic is perhaps the most popular method to assess the extent of statistical heterogeneity within meta-analyses, mostly due to its uncritical promotion within the Cochrane Collaboration.2 The study by Deppen et al demonstrates an overreliance on the I2 statistic and its interpretation that can sometimes be misleading.3 The value of the I2 statistic is not only a product of between-study variability, it is also influenced by the precision of study estimates,3 which, for proportions, is a function of sample size and the number of events of interest. If the actual proportions in the individual studies and the between-study heterogeneity, typically denoted by τ2, are kept constant, but sample sizes are increased, then the I2 will increase. A quick eyeball test for heterogeneity in the study by Deppen et al suggests larger heterogeneity for specificity rather than sensitivity (Figure 2 in the article); however, the I2 statistic (82% for specificity and 87% for sensitivity) suggests otherwise. When the I2 statisticwas first reported,Higgins et al2 suggestedcutoffs todescribeheterogeneityqualitatively (eg,>75% is high). Their benchmark was based on comparative measures (eg, odds ratios), and these behave differently than proportions.More specifically, given equal sample size, the variance of a proportion is smaller than the variance of an odds ratio; therefore, the I2 statistic will tend to be larger for proportions.